"Last week, Mercedes-Benz invited journalists to Los Angeles to test the company's Drive Pilot advanced driver-assist system (ADAS), the first Level 3 system technology to be state-certified in the United States. L.A. traffic didn't disappoint.
As defined by the Society of Automotive Engineers, Level 3 means "conditional automated" operation, such that the driver is legally allowed to take his or her eyes off the road and hands off the wheel. Level 2 ADAS -- incorporating now-commonplace functions such as active lane keeping, dynamic cruise control, collision avoidance and emergency braking -- require drivers to remain eyes-up and hands-on, even if it's only a finger on the wheel.
Available next year for an upfront subscription fee of $2,500, Drive Pilot will have a few requirements of its own. Speed has to be 40 mph or lower, in dry and daylight conditions, with well-marked pavement and a vehicle ahead to follow. For however long these conditions are satisfied, the operator is free to watch video, play games, message, browse or otherwise relax.
When speed exceeds 40 mph the car will drop out of Level 3 operation. Drivers then have to re-initiate the ADAS systems. If everything is jake, indicators will signal that Drive Pilot is again available.
Stipulated, I'm a big dumb monkey, but I had issues. The redundant button-pushing required as the car passed in and out of what the software folks called the Operational Design Domain was highly non-optimal. I'm told the extra clicks to activate and confirm Drive Pilot were necessary to satisfy SAE and European regulations.
Drivers can look down but they can't close their eyes without drawing the biometric scrutiny of the system, which will warn, alert (audible, visual, haptic) and then gradually count down from 10, giving drivers plenty of time to take control. Wake up, sleepy head! The vast majority assume control within about 4 seconds, Mercedes engineers said.
I struggled to keep my paws off the pedals and wheel -- touching either will kick the system out of Level 3. It will also disengage if it loses sight of a lead vehicle. A few times when I was obliged to assume command, I stepped too hard on the accelerator, causing the car to lunge. Oops. Learning curve ahead.
These fretful moments -- known to human-factors engineers as the handoff -- are the most problematic for autonomous driving, when responsibility in an emergent situation would seem to be tossed between driver and car like a hot potato. Mercedes-Benz is explicit on this point: If, during those 10 seconds, the system fails to operate as designed, that's on them. Otherwise, the driver is responsible.
As it often is, Tesla is the elephant in the room. The Silicon Valley car company run by Elon Musk has poured billions into autonomous technology, which began appearing in Tesla cars back in 2014 under the billing of Autopilot. Tesla's Enhanced Autopilot, including automatic route-following navigation, auto park and auto summon functions, is currently a $6,000 option. While more capable than Drive Pilot, the Tesla system is still Level 2, so drivers can't just set and forget. The difference is that Tesla tries to minimize these legally indemnifying check-ins whereas Mercedes-Benz depends on them.
It might be tempting to think of the companies at two points along the same technical trajectory. But Tesla is very much following its own path; a notable example is the company's decision to rely on machine vision and optical cameras to generate the car's virtualized view of the world.
Representing prevailing wisdom, the Mercedes-Benz EQS I drove in Santa Monica was fitted with a mix of sensors operating at multiple wavelengths, including binocular cameras in the windshield and long-range, laser-based LIDAR sensors mounted in the grille. The one on the left is active; the matching unit on the right is decorative, I was told, there to maintain "symmetry." Uh-huh. Clearly, a second LIDAR unit will live there in future models.
Drive Pilot is also capable of automatic route-following, from highway entrance to exit, navigating a pointillistic 3D map of multi-wavelength imagery that tells the car its whereabouts to the centimeter. However, this map is currently limited to California -- Southern California, at that -- and Nevada, the two states where Drive Pilot is approved for public roads.
At the end of the SAE rainbow is Level 5, unrestricted full driving automation. Robotaxis. This is Tesla's play. The latest version of the company's Full Self Driving technology, Version 12 (FSD V12), represents a heroic break from the regime of classical computing. The brain in the Mercedes, for example, is conventionally heuristic, sifting through its sensor-fused environment with vast, pattern-matching algorithms and responding as programmed, if/then. Is the object ahead a tree, a traffic cone, a cyclist?
But there will always be edge cases, instances of stochastic weirdness when patterns don't match and rules are more observed in the breach. In February, for example, Tesla was obliged to recall almost a quarter-million cars beta-testing FSD after NHTSA identified "certain driving maneuvers [that] could potentially infringe upon local traffic laws or customs, which could increase the risk of a collision if the driver does not intervene." Tesla issued the responsive software update in March.
Inevitably, fully autonomous cars of the future will be called upon to respond, to intuit, to improvise, to behave like human beings -- who are, by the way, strangely awesome at driving.
To cross this uncanny valley of autonomy, Tesla has moved to a deeper kind of processing based on generative artificial intelligence. FSD V12 (using the fourth-generation hardware, HW4) effectively abandons the bulk code front-loaded into previous versions, even a map, relying instead on a series of neural nets daisy-chained together, tasked to learn and mimic human driving behavior by watching video clips. FSD V12 was never told to stop at stop signs; it just knows how humans respond to them, and does that.
Initially trained on video clips and telemetry curated from millions of Teslas already on the road, FSD V12 will continue to observe and learn. More data is always better. To that end Tesla has built Dojo, a supercomputing neural-network trainer. Dojo is designed to turn millions of terabytes of video data, gathered from hundreds of millions of driven miles, into something like instinct, reflex, wisdom and experience, to be imparted to succeeding generations of self-driving cars.
Good luck with that. Seriously." [1]
1. OFF DUTY --- Gear & Gadgets -- Rumble Seat: Mercedes-Benz Tests The Autonomous Driving Waters. Neil, Dan. Wall Street Journal, Eastern edition; New York, N.Y.. 30 Sep 2023: D.15.
Komentarų nėra:
Rašyti komentarą